When a process has several control loops sharing the same equipment, moving one valve almost always disturbs more than one measurement - the loops interact. The relative gain array, or RGA, is the classic tool for deciding which manipulated variable should be paired with which controlled variable so that interaction hurts as little as possible. It reduces a messy web of cross-effects to a simple table of numbers you can read at a glance. This page explains what RGA values mean, how values near one flag good pairings while negative or very large values warn of trouble, and how the RGA points toward decoupling or model predictive control when the loops are too tangled to run independently.
Relative Gain Array (RGA) in one line: The relative gain array, introduced by Bristol, is a matrix that measures how much interaction there is between the manipulated and controlled variables of a multivariable process. Each element compares a loop's gain with all other loops open against its gain with all other loops closed. Elements near 1 indicate a pairing with little interaction and good behavior, while values that are negative, near zero, or very large warn of poor pairings, sluggishness, or instability.
The RGA answers a specific question for every possible input-output pairing: how much does the gain from a given manipulated variable to a given controlled variable change when the other loops go from open to closed? If the other loops being in automatic barely affects that gain, the pairing is nearly independent of them, and the RGA element for that pairing is close to one. If closing the other loops dramatically changes or reverses the gain, the pairing is strongly entangled with them, and the element moves far from one.
Each element is a ratio - the open-loop gain over the closed-loop-elsewhere gain - so it carries a clear interpretation without units. The rows correspond to controlled variables and the columns to manipulated variables, and a useful property is that every row and every column sums to one. That constraint means you cannot improve one pairing's number without affecting others, which is precisely why the array, rather than any single number, is what you study.
A great virtue of the RGA is that it needs only steady-state gains, which you can get from a model or from simple step tests, not full dynamic models. That makes it a cheap first screen: before investing in dynamic modeling or advanced control, you can compute the RGA from readily available information and learn a great deal about whether the loops can be run as independent single loops at all.
The reading rules are refreshingly concrete. An element equal to one is ideal - that pairing behaves the same whether or not the other loops are closed, so it can be tuned as if it were on its own. Elements between about zero and one but comfortably away from the extremes indicate manageable interaction. The strategy is to pair variables on elements near one and to avoid pairing on elements far from it, choosing one element per row and column so that no controlled variable and no manipulated variable is used twice.
The warning signs are just as clear. A negative RGA element is a red flag: it means the gain of that loop reverses sign depending on whether the other loops are in automatic or manual. Pair on a negative element and the loop can be stable with the others in manual and unstable with them in automatic, or vice versa - a fragile, dangerous arrangement that should be avoided. Very large positive elements, far above one, warn of severe interaction and extreme sensitivity: the paired loop is only weakly effective on its own and depends heavily on the others, making the whole system touchy and hard to tune. Elements near zero indicate the pairing has almost no effect and should not be chosen.
Putting these together, the RGA gives a ranking. You seek a set of pairings whose RGA elements are all as close to one as possible, positive, and away from zero. When no such set exists - when the best available pairings still leave large or negative elements - the array is telling you something important: the process is too interactive to be controlled well by independent single loops, no matter how you pair them.
When the RGA reveals interaction too severe for clean pairing, it becomes a signpost toward more capable control. One route is decoupling: adding compensators that cancel the effect each loop has on the others, letting the remaining loops be tuned as if independent. The RGA helps here too, indicating how strong the coupling is and therefore how much a decoupler must correct. The other route, favored when interaction stacks up with dead time and constraints, is model predictive control, which coordinates all the manipulated and controlled variables together instead of pretending they are separate. The RGA does not design either solution, but it justifies moving to one.
It is worth being clear about the RGA's limits. Because the standard array uses only steady-state gains, it can miss interactions that only appear in the dynamics - two loops that look decoupled at steady state may still fight each other transiently. Dynamic and frequency-dependent extensions of the RGA exist for that reason. The steady-state RGA remains the everyday screening tool precisely because it is cheap and usually points in the right direction, but a negative or large element deserves confirmation before you commit an expensive control strategy to it.
In field operations, the practical inputs and evidence for an RGA analysis live in the historian. Computing the array requires steady-state gains, and the cleanest way to get them is from bump tests - stepping each manipulated variable and reading the settled change in each controlled variable. A cloud SCADA platform such as Merobix, which historizes every manipulated and controlled variable across a site, is where those step responses are captured and read. The same trends also reveal interaction after the fact: two loops that visibly disturb each other every time one moves are showing, in the raw data, the coupling the RGA was meant to predict, which lets an engineer confirm on real operating data whether the chosen pairings are holding up.
An RGA element of one means that loop's gain is unchanged whether the other loops are open or closed, so the pairing is effectively independent of the rest of the process. That is the ideal case: you can pair on it and tune the loop as if it stood alone. Pairings with elements as close to one as possible are what an RGA analysis seeks.
A negative element means the loop's gain reverses sign depending on whether the other loops are in automatic or manual. A controller tuned for one condition can become unstable in the other, so the arrangement is fragile - it may work in commissioning and fail in normal operation, or the reverse. Pairings on negative elements should be avoided; a negative RGA is a strong signal to repair the pairing or move to decoupling or MPC.
The standard RGA uses only steady-state gains, so it does not capture dynamic interactions or dead time; two loops that look decoupled at steady state can still interact transiently. That is why it is a screening tool rather than a final answer. Frequency-dependent and dynamic RGA variants exist to address this, but the steady-state array remains popular because it is cheap and usually points the analysis in the right direction.
Merobix reads your field devices into a cloud SCADA - the real thing behind these terms, live in days from any browser.